decide on topic for last day

[jaberg]

I put an idea about talking about Model Selection / Hyper-parameter optimization on the last day, but basically it was for lack of anything else.

Other ideas?

[jaberg]

Could look at temporal learning algorithms - what do you have for timeseries prediction in Torch?

[clementfarabet]

Yep we could do something temporal. We have the basic ingredients to build recursive neural nets and the likes, but I don't have much experience with it. It could be fun though, more fun than hyper-parameter selection I guess. Do you know a dataset that could be interesting? Or we could just grab a bunch of stock data, and try to do some prediction.

[jaberg]

Yeah hyperparameters are interesting to me, but no fun to explore in a
hands-on session because there are a lot of tools involved and the
interesting experiments take too long.

On the Theano side, a temporal model example would have to be combined
with a little under-the-hood intro to Theano's API, which I was
planning on putting off until the last day anyway. So that's actually
shaping up in my mind: Theano + temporal modeling with theano.scan

On Sun, Jul 1, 2012 at 3:40 AM, Clement Farabet
[email protected]
wrote:

Yep we could do something temporal. We have the basic ingredients to build recursive neural nets and the likes, but I don't have much experience with it. It could be fun though, more fun than hyper-parameter selection I guess. Do you know a dataset that could be interesting? Or we could just grab a bunch of stock data, and try to do some prediction.

---

Reply to this email directly or view it on GitHub:
#3 (comment)

[clementfarabet]

Yeah hyperparameters are interesting to me, but no fun to explore in a
hands-on session because there are a lot of tools involved and the
interesting experiments take too long.

Exactly.

On the Theano side, a temporal model example would have to be combined
with a little under-the-hood intro to Theano's API, which I was
planning on putting off until the last day anyway. So that's actually
shaping up in my mind: Theano + temporal modeling with theano.scan

OK, same on my side, it involves a little bit of hacking, but nothing too crazy.

Let me try to work something out on that side.